2020-11-30 20:10:26

by Stephen Brennan

[permalink] [raw]
Subject: [PATCH] proc: Allow pid_revalidate() during LOOKUP_RCU

The pid_revalidate() function requires dropping from RCU into REF lookup
mode. When many threads are resolving paths within /proc in parallel,
this can result in heavy spinlock contention as each thread tries to
grab a reference to the /proc dentry (and drop it shortly thereafter).

Allow the pid_revalidate() function to execute under LOOKUP_RCU. When
updates must be made to the inode due to the owning task performing
setuid(), drop out of RCU and into REF mode.

Signed-off-by: Stephen Brennan <[email protected]>
---

I'd like to use this patch as an RFC on this approach for reducing spinlock
contention during many parallel path lookups in the /proc filesystem. The
contention can be triggered by, for example, running ~100 parallel instances of
"TZ=/etc/localtime ps -fe >/dev/null" on a 100CPU machine. The %sys utilization
in such a case reaches around 90%, and profiles show two code paths with high
utilization:

walk_component
lookup_fast
unlazy_child
legitimize_root
__legitimize_path
lockref_get_not_dead

terminate_walk
dput
dput

By applying this patch, %sys utilization falls to around 60% under the same
workload.

One item I'd like to highlight about this patch is that the
security_task_to_inode() hook is called less frequently as a result. I don't
know whether this is a major concern, which is why I've included security
reviewers as well.

fs/proc/base.c | 50 ++++++++++++++++++++++++++++++++-------------
fs/proc/internal.h | 5 +++++
include/linux/pid.h | 2 ++
kernel/pid.c | 12 +++++++++++
4 files changed, 55 insertions(+), 14 deletions(-)

diff --git a/fs/proc/base.c b/fs/proc/base.c
index ebea9501afb8..038056f94ed0 100644
--- a/fs/proc/base.c
+++ b/fs/proc/base.c
@@ -1813,12 +1813,29 @@ int pid_getattr(const struct path *path, struct kstat *stat,
/*
* Set <pid>/... inode ownership (can change due to setuid(), etc.)
*/
-void pid_update_inode(struct task_struct *task, struct inode *inode)
+static int do_pid_update_inode(struct task_struct *task, struct inode *inode,
+ unsigned int flags)
{
- task_dump_owner(task, inode->i_mode, &inode->i_uid, &inode->i_gid);
+ kuid_t uid;
+ kgid_t gid;
+
+ task_dump_owner(task, inode->i_mode, &uid, &gid);
+ if (uid_eq(uid, inode->i_uid) && gid_eq(gid, inode->i_gid) &&
+ !(inode->i_mode & (S_ISUID | S_ISGID)))
+ return 1;
+ if (flags & LOOKUP_RCU)
+ return -ECHILD;

+ inode->i_uid = uid;
+ inode->i_gid = gid;
inode->i_mode &= ~(S_ISUID | S_ISGID);
security_task_to_inode(task, inode);
+ return 1;
+}
+
+void pid_update_inode(struct task_struct *task, struct inode *inode)
+{
+ do_pid_update_inode(task, inode, 0);
}

/*
@@ -1830,19 +1847,24 @@ static int pid_revalidate(struct dentry *dentry, unsigned int flags)
{
struct inode *inode;
struct task_struct *task;
-
- if (flags & LOOKUP_RCU)
- return -ECHILD;
-
- inode = d_inode(dentry);
- task = get_proc_task(inode);
-
- if (task) {
- pid_update_inode(task, inode);
- put_task_struct(task);
- return 1;
+ int rv = 0;
+
+ if (flags & LOOKUP_RCU) {
+ inode = d_inode_rcu(dentry);
+ task = get_proc_task_rcu(inode);
+ if (task) {
+ rv = do_pid_update_inode(task, inode, flags);
+ put_task_struct_rcu_user(task);
+ }
+ } else {
+ inode = d_inode(dentry);
+ task = get_proc_task(inode);
+ if (task) {
+ rv = do_pid_update_inode(task, inode, flags);
+ put_task_struct(task);
+ }
}
- return 0;
+ return rv;
}

static inline bool proc_inode_is_dead(struct inode *inode)
diff --git a/fs/proc/internal.h b/fs/proc/internal.h
index cd0c8d5ce9a1..aa6df65ad3eb 100644
--- a/fs/proc/internal.h
+++ b/fs/proc/internal.h
@@ -121,6 +121,11 @@ static inline struct task_struct *get_proc_task(const struct inode *inode)
return get_pid_task(proc_pid(inode), PIDTYPE_PID);
}

+static inline struct task_struct *get_proc_task_rcu(const struct inode *inode)
+{
+ return get_pid_task_rcu_user(proc_pid(inode), PIDTYPE_PID);
+}
+
void task_dump_owner(struct task_struct *task, umode_t mode,
kuid_t *ruid, kgid_t *rgid);

diff --git a/include/linux/pid.h b/include/linux/pid.h
index 9645b1194c98..0b2c54f85e6d 100644
--- a/include/linux/pid.h
+++ b/include/linux/pid.h
@@ -86,6 +86,8 @@ static inline struct pid *get_pid(struct pid *pid)
extern void put_pid(struct pid *pid);
extern struct task_struct *pid_task(struct pid *pid, enum pid_type);
extern struct task_struct *get_pid_task(struct pid *pid, enum pid_type);
+extern struct task_struct *get_pid_task_rcu_user(struct pid *pid,
+ enum pid_type type);

extern struct pid *get_task_pid(struct task_struct *task, enum pid_type type);

diff --git a/kernel/pid.c b/kernel/pid.c
index 0a9f2e437217..05acbd15cfa6 100644
--- a/kernel/pid.c
+++ b/kernel/pid.c
@@ -390,6 +390,18 @@ struct task_struct *get_pid_task(struct pid *pid, enum pid_type type)
}
EXPORT_SYMBOL_GPL(get_pid_task);

+struct task_struct *get_pid_task_rcu_user(struct pid *pid, enum pid_type type)
+{
+ struct task_struct *task;
+
+ task = pid_task(pid, type);
+ if (task && refcount_inc_not_zero(&task->rcu_users))
+ return task;
+
+ return NULL;
+}
+EXPORT_SYMBOL_GPL(get_pid_task_rcu_user);
+
struct pid *find_get_pid(pid_t nr)
{
struct pid *pid;
--
2.25.1


2020-11-30 20:50:32

by Eric W. Biederman

[permalink] [raw]
Subject: Re: [PATCH] proc: Allow pid_revalidate() during LOOKUP_RCU

Stephen Brennan <[email protected]> writes:

> The pid_revalidate() function requires dropping from RCU into REF lookup
> mode. When many threads are resolving paths within /proc in parallel,
> this can result in heavy spinlock contention as each thread tries to
> grab a reference to the /proc dentry (and drop it shortly thereafter).
>
> Allow the pid_revalidate() function to execute under LOOKUP_RCU. When
> updates must be made to the inode due to the owning task performing
> setuid(), drop out of RCU and into REF mode.

So rather than get_task_rcu_user. I think what we want is a function
that verifies task->rcu_users > 0.

Which frankly is just "pid_task(proc_pid(inode), PIDTYPE_PID)".

Which is something that we can do unconditionally in pid_revalidate.

Skipping the update of the inode is probably the only thing that needs
to be skipped.

It looks like the code can safely rely on the the security_task_to_inode
in proc_pid_make_inode and remove the security_task_to_inode in
pid_update_inode.


> Signed-off-by: Stephen Brennan <[email protected]>
> ---
>
> I'd like to use this patch as an RFC on this approach for reducing spinlock
> contention during many parallel path lookups in the /proc filesystem. The
> contention can be triggered by, for example, running ~100 parallel instances of
> "TZ=/etc/localtime ps -fe >/dev/null" on a 100CPU machine. The %sys utilization
> in such a case reaches around 90%, and profiles show two code paths with high
> utilization:

Do you have a real world work-load that is behaves something like this
micro benchmark? I am just curious how severe the problem you are
trying to solve is.

>
> walk_component
> lookup_fast
> unlazy_child
> legitimize_root
> __legitimize_path
> lockref_get_not_dead
>
> terminate_walk
> dput
> dput
>
> By applying this patch, %sys utilization falls to around 60% under the same
> workload.
>
> One item I'd like to highlight about this patch is that the
> security_task_to_inode() hook is called less frequently as a result. I don't
> know whether this is a major concern, which is why I've included security
> reviewers as well.
>
> fs/proc/base.c | 50 ++++++++++++++++++++++++++++++++-------------
> fs/proc/internal.h | 5 +++++
> include/linux/pid.h | 2 ++
> kernel/pid.c | 12 +++++++++++
> 4 files changed, 55 insertions(+), 14 deletions(-)
>
> diff --git a/fs/proc/base.c b/fs/proc/base.c
> index ebea9501afb8..038056f94ed0 100644
> --- a/fs/proc/base.c
> +++ b/fs/proc/base.c
> @@ -1813,12 +1813,29 @@ int pid_getattr(const struct path *path, struct kstat *stat,
> /*
> * Set <pid>/... inode ownership (can change due to setuid(), etc.)
> */
> -void pid_update_inode(struct task_struct *task, struct inode *inode)
> +static int do_pid_update_inode(struct task_struct *task, struct inode *inode,
> + unsigned int flags)
> {
> - task_dump_owner(task, inode->i_mode, &inode->i_uid, &inode->i_gid);
> + kuid_t uid;
> + kgid_t gid;
> +
> + task_dump_owner(task, inode->i_mode, &uid, &gid);
> + if (uid_eq(uid, inode->i_uid) && gid_eq(gid, inode->i_gid) &&
> + !(inode->i_mode & (S_ISUID | S_ISGID)))
> + return 1;
> + if (flags & LOOKUP_RCU)
> + return -ECHILD;
>
> + inode->i_uid = uid;
> + inode->i_gid = gid;
> inode->i_mode &= ~(S_ISUID | S_ISGID);
> security_task_to_inode(task, inode);
> + return 1;
> +}
> +
> +void pid_update_inode(struct task_struct *task, struct inode *inode)
> +{
> + do_pid_update_inode(task, inode, 0);
> }
>
> /*
> @@ -1830,19 +1847,24 @@ static int pid_revalidate(struct dentry *dentry, unsigned int flags)
> {
> struct inode *inode;
> struct task_struct *task;
> -
> - if (flags & LOOKUP_RCU)
> - return -ECHILD;
> -
> - inode = d_inode(dentry);
> - task = get_proc_task(inode);
> -
> - if (task) {
> - pid_update_inode(task, inode);
> - put_task_struct(task);
> - return 1;
> + int rv = 0;
> +
> + if (flags & LOOKUP_RCU) {
> + inode = d_inode_rcu(dentry);
> + task = get_proc_task_rcu(inode);
> + if (task) {
> + rv = do_pid_update_inode(task, inode, flags);
> + put_task_struct_rcu_user(task);
> + }
> + } else {
> + inode = d_inode(dentry);
> + task = get_proc_task(inode);
> + if (task) {
> + rv = do_pid_update_inode(task, inode, flags);
> + put_task_struct(task);
> + }
> }
> - return 0;
> + return rv;
> }
>
> static inline bool proc_inode_is_dead(struct inode *inode)
> diff --git a/fs/proc/internal.h b/fs/proc/internal.h
> index cd0c8d5ce9a1..aa6df65ad3eb 100644
> --- a/fs/proc/internal.h
> +++ b/fs/proc/internal.h
> @@ -121,6 +121,11 @@ static inline struct task_struct *get_proc_task(const struct inode *inode)
> return get_pid_task(proc_pid(inode), PIDTYPE_PID);
> }
>
> +static inline struct task_struct *get_proc_task_rcu(const struct inode *inode)
> +{
> + return get_pid_task_rcu_user(proc_pid(inode), PIDTYPE_PID);
> +}
> +
> void task_dump_owner(struct task_struct *task, umode_t mode,
> kuid_t *ruid, kgid_t *rgid);
>
> diff --git a/include/linux/pid.h b/include/linux/pid.h
> index 9645b1194c98..0b2c54f85e6d 100644
> --- a/include/linux/pid.h
> +++ b/include/linux/pid.h
> @@ -86,6 +86,8 @@ static inline struct pid *get_pid(struct pid *pid)
> extern void put_pid(struct pid *pid);
> extern struct task_struct *pid_task(struct pid *pid, enum pid_type);
> extern struct task_struct *get_pid_task(struct pid *pid, enum pid_type);
> +extern struct task_struct *get_pid_task_rcu_user(struct pid *pid,
> + enum pid_type type);
>
> extern struct pid *get_task_pid(struct task_struct *task, enum pid_type type);
>
> diff --git a/kernel/pid.c b/kernel/pid.c
> index 0a9f2e437217..05acbd15cfa6 100644
> --- a/kernel/pid.c
> +++ b/kernel/pid.c
> @@ -390,6 +390,18 @@ struct task_struct *get_pid_task(struct pid *pid, enum pid_type type)
> }
> EXPORT_SYMBOL_GPL(get_pid_task);
>
> +struct task_struct *get_pid_task_rcu_user(struct pid *pid, enum pid_type type)
> +{
> + struct task_struct *task;
> +
> + task = pid_task(pid, type);
> + if (task && refcount_inc_not_zero(&task->rcu_users))
> + return task;
> +
> + return NULL;
> +}
> +EXPORT_SYMBOL_GPL(get_pid_task_rcu_user);
> +
> struct pid *find_get_pid(pid_t nr)
> {
> struct pid *pid;

2020-12-01 23:53:43

by Stephen Brennan

[permalink] [raw]
Subject: Re: [PATCH] proc: Allow pid_revalidate() during LOOKUP_RCU

[email protected] (Eric W. Biederman) writes:
> Stephen Brennan <[email protected]> writes:
>
>> The pid_revalidate() function requires dropping from RCU into REF lookup
>> mode. When many threads are resolving paths within /proc in parallel,
>> this can result in heavy spinlock contention as each thread tries to
>> grab a reference to the /proc dentry (and drop it shortly thereafter).
>>
>> Allow the pid_revalidate() function to execute under LOOKUP_RCU. When
>> updates must be made to the inode due to the owning task performing
>> setuid(), drop out of RCU and into REF mode.
>
> So rather than get_task_rcu_user. I think what we want is a function
> that verifies task->rcu_users > 0.
>
> Which frankly is just "pid_task(proc_pid(inode), PIDTYPE_PID)".
>
> Which is something that we can do unconditionally in pid_revalidate.
>
> Skipping the update of the inode is probably the only thing that needs
> to be skipped.
>
> It looks like the code can safely rely on the the security_task_to_inode
> in proc_pid_make_inode and remove the security_task_to_inode in
> pid_update_inode.
>

This makes sense, I'll get rid of the get_task_rcu_user() stuff in a v2.

>
>> Signed-off-by: Stephen Brennan <[email protected]>
>> ---
>>
>> I'd like to use this patch as an RFC on this approach for reducing spinlock
>> contention during many parallel path lookups in the /proc filesystem. The
>> contention can be triggered by, for example, running ~100 parallel instances of
>> "TZ=/etc/localtime ps -fe >/dev/null" on a 100CPU machine. The %sys utilization
>> in such a case reaches around 90%, and profiles show two code paths with high
>> utilization:
>
> Do you have a real world work-load that is behaves something like this
> micro benchmark? I am just curious how severe the problem you are
> trying to solve is.
>

We have seen this issue occur internally with monitoring scripts
(perhaps a bit misconfigured, I'll admit). However I don't have an exact
sample workload that I can give you.

Thanks,
Stephen